Download Atlas Local
All of Atlas with the AI running entirely on your own machine — chat, voice, and search, no cloud round-trip. Activate once with your license key, then use it completely offline.
Requires an NVIDIA GPU on Windows/Linux, or Apple Silicon on macOS.
Linux with an RTX 20-series GPU (2060 / 2070 / 2080) or a GTX 1660? Download the Turing build — the standard Linux download requires an RTX 30-series or newer.
Existing cloud subscribers
On a legacy Atlas Cloud plan? Download the cloud build below — your installers are unchanged.
System Requirements
Atlas Local runs its AI models on your own hardware, so the language model needs a supported GPU. Check your machine below before you buy.
Windows
- OS: Windows 10 or later
- GPU: NVIDIA GPU with CUDA — required to run the local AI model
- VRAM: ~10–12 GB for the default Mistral Nemo 12B; 8 GB → choose Qwen3 4B
- RAM: 16 GB system memory recommended
- Disk: ~3 GB first-run download (~8 GB with the default 12B model), plus your vault + search index
macOS
- OS: macOS 12 (Monterey) or later
- Chip: Apple Silicon (M1 or later) — runs on MLX/Metal
- Unified memory: 16 GB → Qwen3 4B; 24–32 GB → the default Mistral Nemo 12B
- Disk: ~3 GB first-run download (~8 GB with the default 12B model), plus your vault + search index
Linux
- OS: Ubuntu 22.04+ / Fedora 38+
- GPU: NVIDIA GPU with CUDA — required to run the local AI model. RTX 30-series or newer uses the standard download; RTX 20-series and GTX 1660 need the Turing build linked above
- VRAM: ~10–12 GB for the default Mistral Nemo 12B. On 8 GB cards Atlas picks Qwen3 4B for you at first launch
- RAM: 16 GB system memory recommended
- Disk: ~3 GB first-run download (~8 GB with the default 12B model), plus your vault + search index
Models & first-run downloads
On first launch Atlas downloads an essential set of about 3 GB: the smallest language model (Qwen3 4B), Whisper Base speech-to-text, and Kokoro text-to-speech. You can switch language models any time in Settings → AI.
- Mistral Nemo 12B — ~7.5 GB (default on Windows/Linux)
- Qwen3 4B / 8B / 14B — ~2.5 / 5 / 9 GB
- Ministral 3 14B — ~8.2 GB, 256K-token context
- Speech-to-text: Whisper Base — ~290 MB (English)
- Text-to-speech: Kokoro — ~350 MB
Good to know
- GPU is required for chat. The local language model runs on your GPU (NVIDIA CUDA on Windows/Linux, Apple Silicon on macOS). Search and voice components use ONNX and fall back to CPU when needed, but the chat model does not run on CPU.
- Optional system tools: espeak-ng and ffmpeg sharpen voice pronunciation and audio handling. A guided install lives in Settings → Voice — neither is required to get started.
Real-world performance
Atlas runs the model on your machine, so speed depends on your hardware. These are measured medians from people who opted into anonymous performance reporting — not our benchmarks on our machines.
202 responses · 11 installs · last 90 days · Atlas 2.3.0 · updated 2026-08-08
Generation speed
best result per machine · tokens per second · higher is betterTime to first token
How long before it starts replying · lower is better| Hardware | Model | Atlas | tok/s | p95 | First token |
|---|---|---|---|---|---|
| RTX 3090 · 24 GBLinux | Qwen3 4Bq4_k_m | 2.3.0 | 86.1 | 105.3 | 1.41 s |
| RTX 3090 · 24 GBLinux | Mistral Nemo 12Bq4_k_m | 2.3.0 | 67.2 | 74.6 | 8.63 s |
| RTX 3090 · 24 GBLinux | Ministral 3 14Bq4_k_m | 2.3.0 | 56.7 | 58.0 | 1.47 s |
| RTX 3090 · 24 GBLinux · thinking mode | Qwen3 14Bq4_k_m | 2.3.0 | 55.3 | 58.1 | 6.21 s |
| Apple M4 Pro · 48 GB unifiedmacOS | Qwen3 4B4bit | 2.3.0 | 54.5 | 54.6 | 1.62 s |
| RTX 3090 · 24 GBLinux | Qwen3 14Bq4_k_m | 2.3.0 | 52.5 | 61.8 | 0.78 s |
| Apple M4 Pro · 48 GB unifiedmacOS · thinking mode | Qwen3 4B4bit | 2.3.0 | 48.6 | 54.0 | 12.50 s |
| Apple M4 Pro · 48 GB unifiedmacOS · thinking mode | Qwen3 14B4bit | 2.3.0 | 23.7 | 27.4 | 17.84 s |
| Apple M4 · 16 GB unifiedmacOS | Qwen3 4B4bit | 2.3.0 | 23.7 | 37.6 | 3.52 s |
| Apple M3 · 24 GB unifiedmacOS · thinking mode | Qwen3 4B4bit | 2.3.0 | 16.9 | 34.8 | 36.30 s |
| Apple M3 · 24 GB unifiedmacOS | Qwen3 4B4bit | 2.3.0 | 9.1 | 35.0 | 4.40 s |
Medians across every reported response, so half of real replies are faster and half slower. Each bar is one machine at its best measured configuration, named beneath it — the table lists every model that machine ran. Time to first token is recorded only for streamed replies, and excludes the first response after a model loads, which carries one-time warmup. Your own speed will vary with model size, context length, and how much of your vault the answer needs. Memory is dedicated graphics memory on a discrete card, and the machine’s total unified memory on Apple Silicon, where one pool is shared between the system and the model. Atlas Local is new, so these figures come from 11 installs in total — treat them as early signal, not a settled benchmark. How this data is collected.
Getting Started After Download
- 1
Install Atlas
Run the installer for your platform. On Windows, double-click the .exe file. On macOS, open the .dmg and drag Atlas to Applications. On Linux, make the .AppImage executable or install the .deb package.
- 2
Create or select a vault folder
When you first launch Atlas, choose a folder on your machine to store your notes. This is your vault—all your notes, files, and personal knowledge will live here as plain Markdown files. You can use an existing folder with notes or start fresh.
- 3
Write your first note
Click the Files tab and create a new note. Write some thoughts, ideas, or project notes in Markdown format. Your notes are immediately saved as files in your vault folder.
- 4
Try asking Atlas a question about your notes
Go to the Chat tab and ask Atlas something related to what you wrote. The AI will use local semantic search to find relevant context from your vault and provide an answer grounded in your own notes. Everything runs on your own machine—no cloud round-trip.
- 5
Explore voice input, CRM, and calendar features
Try voice input (push-to-talk or wake word mode), add people and organizations to your CRM, connect your Google Calendar, and explore the knowledge graph to visualize connections between your notes. See the documentation for detailed guides.
Troubleshooting
Atlas won't start on Linux
Ensure libwebkit2gtk-4.0-37 is installed. On Ubuntu/Debian: sudo apt install libwebkit2gtk-4.0-37. On Fedora: sudo dnf install webkit2gtk3. For .AppImage files, make them executable: chmod +x Atlas-*.AppImage.
Atlas opens on Linux, but chat fails with a CUDA error
If the first message fails with no kernel image is available for execution on the device, you have the wrong build for your GPU. Atlas's GPU code is compiled for one NVIDIA architecture generation per download: the standard Linux build targets RTX 30-series and newer, and the Turing build targets the RTX 20-series and GTX 1660. Install the other one — the app logs which build it is and which GPU it found at startup. Pascal cards (GTX 1060 / 1080) and the RTX 50-series have no GPU build yet.
Where are my notes stored?
Your notes are stored in the vault folder you selected during setup—plain Markdown files on your machine. You can open this folder anytime from Atlas or navigate to it in your file explorer. You have complete ownership of your data.
macOS says Atlas is from an unidentified developer
Right-click the Atlas app in Applications, select "Open", then click "Open" in the dialog. This adds an exception for Atlas. You only need to do this once. Alternatively, you can go to System Preferences → Security & Privacy → General and click "Open Anyway" after the first launch attempt.
AI features aren't working
Atlas Local runs all AI on your own machine. If chat or voice isn't responding, check that you've activated your license (Settings → License & About), that the model finished downloading on first launch, and that your hardware meets the requirements (NVIDIA/CUDA on Windows/Linux, Apple Silicon on macOS). See the documentation for step-by-step troubleshooting.
Atlas Local — where do I find my license key?
Sign in and open your dashboard — your key is shown in the "Atlas Local License" section. Paste it into Atlas Local on first launch to activate; after that the app runs fully offline. Don't have a license yet? Get Atlas Local or use the "Get Atlas Local" button in your dashboard.
Does Atlas Local need a GPU?
Yes — Atlas Local runs all AI on your own hardware. On Windows and Linux that means an NVIDIA (CUDA) GPU; on macOS, an Apple Silicon (M1 or later) Mac. The app still opens on unsupported hardware, but AI chat and voice will be unavailable or very slow.
For more help, visit our documentation or reach out at info@atlasnotes.io.